It is becoming f*cking dangerous
A team of devs from China just made the same agent task cost $0.20 instead of $10, and they explained how in one line of their release notes:
"Stronger benchmark performance, faster task completion, and shorter outputs than the previous version."
The numbers on DeepSeek's own DeepSWE chart:
- Flash-1017: about $0.20 per task
- Opus 5 at its top setting: around $10
- that's 50 tasks for the price of one
The trick is the part everyone skips: "shorter outputs". An agent pays for every token it writes. Fewer words per step = a cheaper finished task, not just a cheaper token
So the move isn't replacing your best model. Flash does the long, wordy steps, Opus 5.5 only reads the final review
A dev team would burn a monthly budget on this. You can run it with $5
Full Claude Code + DeepSeek breakdown below